Deep learning based harmful algal blooms modeling in inland water using multimodal monitoring

Abstract

Harmful algal blooms (HABs) represent a growing environmental and public health concern in inland water systems, which are driven by complex interactions among hydrological, climatic, and anthropogenic processes. Conventional in-situ monitoring and numerical modeling approaches have provided valuable insights into algal management. However, they have faced critical limitations in capturing the rapid, spatially heterogeneous, and event-driven phenomena of algal proliferation in dynamic freshwater environments. Therefore, an advanced monitoring framework is required that could deal with large-scale, high-frequency, and multi-source observations to effectively present the spatiotemporal dynamics of HABs. This dissertation presents a comprehensive deep learning framework for monitoring, modeling, and water management by integrating multi-source remote sensing, in-situ observation, and environmental datasets within an artificial intelligence (AI) paradigm. First, a deep learning-based super-resolution (SR) approach was developed to enhance the spatial resolution of satellite imagery in inland waters. Using convolutional neural networks and generative adversarial networks, satellite images were reconstructed at sub-pixel precision, providing fine-resolution chlorophyll-a (Chl-a) distribution maps from the Sentinel-2 dataset. The findings of the first research could provide advancement bridges with scale differences between coarse satellite imagery and narrow inland water bodies, thereby the deep learning-based SR approach contributed to improving the spatial representativeness of algal bloom monitoring. Second, a probabilistic machine learning frameworks were implemented to directly simulate phytoplankton abundance from the hyperspectral remote sensing data. Bayesian neural network (BNN) and natural gradient boosting (NGBoost) were utilized to estimate algal cell concentration and to quantify predictive uncertainty according to the measurement noise and data scarcity. These results demonstrated that probabilistic models not only showed superior performance but also provided a credible measure of monitoring uncertainty, which could be essential for ecological interpretation and risk-aware bloom forecasting. Third, a multi-modal deep learning was developed for algal phyla classification by integrating heterogeneous modalities, including microscopic images and particle properties. The proposed multimodal learning combined visual and quantitative representations of the algal phyla from data fusion strategies. The framework achieved high classification performance across major algal phyla and was further interpreted using eXplainable AI (XAI) using Shapley Additive Explainations (SHAP), and Gradient- weighted Class Activation Mapping (Grad-CAM). Therefore, this study demonstrated that multimodal learning could capture and integrate the morphological and particle-based features that contributed to the differentiation for algal identification. Fourth, a tower-based hyperspectral monitoring system was integrated with a multimodal deep learning model to predict high-frequency monitoring of cyanobacterial blooms. This research incorporated continuous hyperspectral reflectance, environmental, and in-situ RGB imagery to simulate temporally bloom dynamics. The integration of high-frequency tower observations with multimodal deep learning allows continuous detection of short-term variations in bloom intensity and spatial heterogeneity, providing a practical framework for real-time HAB monitoring in inland waters. Hence, this dissertation advances HAB research by integrating in-situ observations, remote sensing, and data-driven modeling into a reliable and complementary framework. The findings of this dissertation deal with scalable approaches that combine spatial precision, temporal continuity, and model interpretability. The proposed AI-driven framework can enhance transparency and reliability in ecological prediction. This research could contribute to the development of a water quality monitoring system and sustainable water resource management in response to accelerating climatic and anthropogenic pressure within AI paradigm.